Websites already have hidden APIs that agents could use instead of clicking through pages — 3.6 times faster

Researchers found that every modern website exposes internal APIs (the same endpoints the site's own code uses) that AI agents could call directly instead of simulating a browser and clicking through the visible interface. This means agents could fetch information from websites 3 to 5 times faster, and the system they built automatically learns which endpoints work on each site by watching real traffic, then shares them in a searchable index.

Right now, AI agents interact with websites the way humans do — they load pages, parse the HTML, click buttons, wait for responses. This is slow and fragile because the website's human interface is designed for eyeballs, not machines. The structural change is that agents could skip the performance penalty of the human interface entirely and just call the API endpoints the website already built for itself. This means the bottleneck that has slowed down every web-using agent for the last few years might be technically solvable without waiting for websites to redesign. The catch is adoption — websites have no incentive to help agents unless they benefit somehow, so the system proposes a payment model to align that.